1147 A Multimodal Machine-Learning Framework for Quantifying Sleep Health Using Actigraphy and Subjective Measures in People with Stroke

Abstract Introduction Sleep health is a multidimensional construct that spans objective sleep–wake patterns and subjective experiences. Current measures mainly rely on self-report items. Few approaches have integrated objective and subjective indicators into a unified, data-driven measure of sleep health. The purpose of this study was to develop and validate a multimodal, machine-learning–based Sleep Health Index (SHI) that integrates high-resolution actigraphy data with sleep related self-report measures. Methods Data were collected at 10-, 60-, and 90-days post stroke. Participants wore an actigraph for 7 days and completed the Epworth Sleepiness Scale (ESS) and Insomnia Severity Index (ISI) at each time point. Sleep regularity, timing, fragmentation, efficiency, and circadian stability were extracted from multi-day granular actigraphy recordings and derive latent actigraphy features using a temporal convolutional autoencoder. Subjective indicators included sleep satisfaction and daytime alertness. A hierarchical multimodal latent variable model jointly characterized objective and subjective domains, forming the basis of an integrative sleep health construct. The SHI was generated using supervised machine-learning models trained to predict multimodal latent factor scores, with construct validity assessed via associations with subdomains of RuSATED. Results Data from 115 participants was used. Preliminary associations showed coherent convergence across domains. RuSATED defined sleep efficiency aligned strongly with objective sleep consolidation (r = 0.67 for efficiency; r = −0.65 for WASO; RuSATED defined timing was inversely related to variability in mid-sleep timing (r = −0.44, p 0.001). Daytime alertness was associated with wake-time regularity (r = −0.47, p 0.001), and sleep dissatisfaction correlated with greater WASO (r = 0.29, p = 0.04). Conclusion This study demonstrates a novel framework for quantifying sleep health that combines high-resolution actigraphy with subjective experiences using multimodal latent modeling and machine learning. The resulting SHI provides a comprehensive, sensitive, and data-driven characterization of sleep health. This approach may apply across diverse clinical populations and offers a flexible, scalable method for integrating wearable data and self-report into a single interpretable construct of sleep health. Support (if any) This research was supported by the NIH NINR, Award Number R01NR018979.